CN104573716A - Eye fundus image arteriovenous retinal blood vessel classification method based on breadth first-search algorithm - Google Patents

Abstract

The invention discloses a method for classifying arteriovenous retinal blood vessels of a fundus image based on a breadth search algorithm, comprising: (1) acquiring a global blood vessel set and optic disc positioning information of the fundus image, and the global blood vessel set is the fundus image in the fundus image A collection of all blood vessels, the optic disc positioning information includes the center of the optic disc of the fundus image; (2) determining main vessels according to the global vessel set and optic disc positioning information, and classifying the main vessels to obtain main vessel classification information; (3) Using the main vessel classification information to classify the blood vessels in the global blood vessel collection using a SAT-based breadth search algorithm to obtain global classification information. The invention first obtains the classification information of the main blood vessels around the optic disc, and expands and diffuses from the main blood vessels based on the SAT breadth search algorithm to obtain all the blood vessels, realizing a complete automatic blood vessel classification method without manual intervention and high classification accuracy.

Description

Arteriovenous Retinal Vessel Classification Method for Fundus Images Based on Breadth Search Algorithm

technical field

The invention relates to the technical field of computer-aided diagnosis, in particular to a method for classifying arteriovenous retinal blood vessels of fundus images based on a breadth search algorithm.

Background technique

With the rapid development of the field of artificial intelligence in computer technology, computer-aided diagnosis technology is also gradually developed. Computer-aided diagnosis technology refers to the use of imaging, medical image processing technology and other possible physiological and biochemical means, combined with computer analysis and calculation, to assist radiologists in finding lesions and improving the accuracy of diagnosis.

Usually, computer-aided diagnosis in medical imaging is divided into three steps, as follows: the first step is to extract the lesion from the normal structure; the second step is to quantify the image features; the third step is to process the data and draw conclusions .

Because the computer can fully use the image information to carry out accurate quantitative calculations, remove human subjectivity, and avoid "various" diagnostic results caused by differences in personal knowledge and experience; therefore, its results are unambiguous and deterministic. It makes the diagnosis more accurate and more scientific.

With the development of modern high-tech, computer-aided diagnosis will be integrated with image processing and PACS systems, making it easier to operate and more accurate, and its clinical application will be further expanded.

In medical testing, the eye is the only organ that can be inspected non-destructively and is rich in information. Studies have pointed out that in retinal vascular diseases, narrowing of blood vessels, diffuse narrowing, arteriovenous cross compression, changes in blood vessel course, copper wire arteries, hemorrhage, cotton wool spots, hard exudates, and retinal nerve fiber layer defects are significantly related to stroke. sex. And for the prediction of stroke, the fundus examination only costs 40 yuan, while the MRI examination costs thousands of yuan, and the carotid artery ultrasound also costs 140 yuan. In contrast, fundus examination is the most cost-effective. The fully automated method of computer analysis of fundus images, including the ability to provide immediate classification of retinal lesions without expert advice, and the establishment of a system for predicting three high complications based on fundus blood vessels and optic nerves has real economic significance. Therefore, the detection of retinal vascular lesions plays a prominent role in the auxiliary detection of stroke. Among them, the construction of an automatic detection system for retinal vascular lesions caused by arteriovenous cross compression is a key part.

Blood vessel segmentation, optic disc positioning and vessel classification (arteriovenous split) on fundus images are the basis of retinal vessel lesion detection. The existing blood vessel segmentation methods need to manually add labeling information, and the degree of automation is not high.

Contents of the invention

Aiming at the deficiencies of the prior art, the present invention provides a method for classifying arteriovenous retinal blood vessels of fundus images based on a breadth search algorithm.

A method for classifying arteriovenous retinal blood vessels of fundus images based on a breadth search algorithm. First, the global blood vessel set and optic disc positioning information of the fundus image are obtained. The global blood vessel set is a collection of all blood vessels in the fundus image, and the The optic disc positioning information includes the optic disc center of the fundus image, and then according to the global blood vessel set and the optic disc positioning information, the blood vessels in the global blood vessel set are classified into arteriovenous retinal vessels, and the following steps are performed during classification:

(1) Determine the main blood vessel according to the global blood vessel set and the optic disc positioning information, and classify the main blood vessels to obtain the main blood vessel classification information;

The main blood vessels around the optic disc are taken because generally speaking, when the blood vessels just originate from the center of the optic disc, the arteries and veins still have some degree of distinction. When blood vessels extend farther away from the optic disc, the degree of discrimination becomes smaller, and even professional doctors can hardly use the local information of blood vessels to classify arteries and veins.

In the present invention, the main blood vessels are determined by the following method:

Take the area extending several pixels from the center of the optic disc as the adjacent area of the optic disc (that is, the area within several pixels from the center of the optic disc as the adjacent area of the optic disc), and the length of the adjacent area of the optic disc is greater than the preset classification length Threshold the vessel as the main vessel.

In the present invention, R pixels are expanded outward, that is, the center of the optic disc is the center, and the area with R as the radius is used as the adjacent area of the optic disc, and the blood vessels in the determined adjacent area of the optic disc whose length is greater than the preset classification length threshold are used as the main blood vessels.

Wherein, the size of the radius R and the classification length threshold is determined according to the size of the fundus picture and the actual situation. Preferably, the value of R is 100-150, and the classification length threshold is 50-65.

After the main vessel is determined, the main vessel classification information is obtained by classifying the main vessel through the following steps:

(1-1) Obtain the average caliber of each main vessel, and designate the main vessel with the largest average caliber as the venous vessel;

From an anatomical principle, the thickest blood vessel in the primary blood vessels (main blood vessels) around the optic disc is generally a venous blood vessel.

Usually the fundus image is a two-dimensional image, which reflects the diameter of the blood vessel in the fundus image and actually the width of the blood vessel in the fundus image.

(1-2) cutting each main vessel into several segments to obtain corresponding main vessel slices;

The use of slices instead of the mean value of the whole blood vessel segment is because it can increase the number of samples and facilitate clustering to distinguish arteries and veins. At the same time, because the length of blood vessels is not uniform, the consistency of feature dimensions can be guaranteed.

(1-3) Extract the eigenvectors of each main vessel slice, and use the clustering method to cluster the main vessel slices into two classes based on the feature vectors, and use the class where the main vessel slice corresponding to the venous vessel is located As a venous vessel, the other as an arterial vessel;

Preferably, the K-means clustering method is used in the present invention to cluster the main blood vessel slices into two categories.

(1-4) For each main vessel, the class in which (the main vessel) has more main vessel slices is used as the classification result of the main vessel.

As preferably, in described step (1-3), extract the feature vector of each main blood vessel slice by following method:

Obtain the color information of all pixels in the region within a few pixels from the center of the main blood vessel, and use the mean value of the color information of all pixels in the region as the feature vector of the main blood vessel slice.

The color information includes the RGB value and the HSL value of the sampling point, and the mean value of the color information of all pixels is used as the feature vector of the main blood vessel slice.

Further preferably, when extracting the feature vectors of each main blood vessel slice, the color information of all pixels in the area whose distance to the center of the blood vessel is less than a preset distance threshold is obtained, and the preset distance threshold is 5 to 8 pixels. That is, the color information of 5 to 8 pixel points is respectively acquired along the vessel center of the main vessel slice to its surroundings.

(2) Using the main vessel classification information to classify the blood vessels in the global vessel collection using a SAT-based breadth search algorithm to obtain global classification information.

Breadth search is performed on the global blood vessel set, and the SAT-based breadth search algorithm uses three constraints to transfer blood vessel categories during the search process: the two blood vessels in the cross are respectively marked as two types of blood vessels; the three blood vessels in the trifurcated structure are marked as ; Among the three vessels in the trifurcated structure, if the sum of the angle between one of the three vessels and the remaining two vessels is less than or equal to 270 degrees, it is judged as the trifurcated structure of constraint 2, that is, the three vessels belong to the same type of vessel, otherwise it is not done determination.

By using the above three constraints, it is possible to better distinguish the situation where the intersection is misjudged as a trifurcation due to the omission of segmentation in the blood vessel segmentation, and the classification accuracy is improved. On the other hand, based on the constraints, the blood vessels with high reliability get the classification results first, and the blood vessels with low reliability make corrections for the areas that the high-confidence blood vessels cannot reach, which can make the global blood vessel labeling credible. The degree is improved, thereby improving the classification effect.

The present invention designates thicker main vessels as venous vessels, so in the entire SAT breadth search algorithm, the thicker the vessel, the higher the reliability of the obtained classification result of the vessel.

Unless otherwise specified, in the present invention, parameters such as length, distance, and picture size are measured in units of pixels.

Compared with the prior art, the present invention first obtains the classification information of the main blood vessels around the optic disc, and expands and diffuses from the main blood vessels based on the SAT breadth search algorithm to obtain all blood vessels, realizing a complete automatic blood vessel classification method without manual work Intervention, and the classification accuracy is high.

Description of drawings

Fig. 1 is the fundus image of the present embodiment;

Fig. 2 is the flowchart of the arteriovenous retina vessel classification based on the fundus image of breadth search algorithm;

Fig. 3 is the flow chart that carries out blood vessel segmentation to fundus image in the present embodiment;

4 is a schematic diagram of an original blood vessel set obtained by blood vessel segmentation;

5 is a schematic diagram of a global vessel set obtained by vessel segmentation;

Fig. 6 is a flow chart of performing optic disc positioning on the fundus image in this embodiment;

FIG. 7 is a schematic diagram of the global classification information of the arteriovenous retinal vessel classification of the fundus image in this embodiment.

Detailed ways

The present invention will be described in detail below with reference to the drawings and specific embodiments.

In this embodiment, the fundus image shown in FIG. 1 is taken as an example to illustrate the method for classifying arteriovenous retinal blood vessels based on the breadth search algorithm. The size of the fundus image is 3000×3000. There are bright rings in the fundus image due to the ring reflection caused by photography, the non-vascular step edge around the optic disc, patchy lesions, and hemorrhagic lesions.

The arteriovenous retinal vessel classification of the fundus image based on the breadth search algorithm is adopted for the fundus image, and the classification process is shown in Figure 2, including the following steps:

(1) Obtain the global blood vessel set (i.e. the final blood vessel set) and the optic disc positioning information of the fundus image, the global blood vessel set is the collection of all blood vessels in the fundus image, and the optic disc positioning information includes the optic disc center of the fundus image;

In this embodiment, the global blood vessel set of the fundus image is obtained by segmenting the blood vessels of the fundus image. The specific process is shown in Figure 3, including the following steps:

(1-1) Carry out wavelet transform (IUWT wavelet) to fundus image, carry out binarization process to the fundus image after wavelet transform according to preset binarization threshold value, and extract the center in the fundus image after binarization process lines and edges, to get the vascular tree;

(1-2) Disconnect the bifurcations of the vascular tree to obtain blood vessel segments, and perform line segmentation on each blood vessel segment to obtain blood vessels, and combine them to obtain the original blood vessel set.

When disconnecting the bifurcation of the vascular tree: when multiple centerlines of the vascular centerlines in the vascular tree converge to one point, remove the central point (the intersection point of convergence) to obtain multiple individual vascular centerlines.

When performing line segmentation on each blood vessel segment: take each central line as a blood vessel segment. The blood vessel segment is a curve, and the traditional method of line segmentation in image processing is used to approximate the curve with multiple straight lines. A plurality of straight lines are obtained, and each straight line represents a blood vessel, and the collection of all straight lines is the original blood vessel set.

(1-3) Determining mis-segmented blood vessels. In this embodiment, the mis-segmented blood vessels are obtained from the first type of mis-segmented blood vessels and the second type of mis-segmented blood vessels, and the first type of mis-segmented blood vessels and the second type of mis-segmented blood vessels are deleted from the original blood vessel collection. blood vessels, the global blood vessel set (ie, the final blood vessel set) is obtained.

For the mis-segmentation caused by ring-shaped reflection, the segmented blood vessels have the structural characteristics of a ring composed of small blood vessels compared with normal blood vessels.

For the mis-segmentation caused by the step edge around the optic disc, the segmented blood vessels have no special characteristics in RGB color space (ie channel) and structure. The mis-segmented blood vessel is composed of the background around the optic disc, because it is close to the optic disc, and the background color around the optic disc is similar to the color of ordinary blood vessels compared to the background far away from the optic disc; structurally, because it exists in isolation, It is also difficult to distinguish from the structure when it is mixed with the blood vessels around the optic disc. If the judgment is forced to be made from the structure, it will easily cause a lot of misjudgment. However, the background on both sides of the blood vessel has a large color difference in the RGB color space, because one side of the background is composed of the optic disc and the other side is composed of a common background. In fact, for common blood vessels, the backgrounds on both sides are composed of common backgrounds or optic discs.

For mis-segmentation caused by patchy lesions and hemorrhagic lesions, the segmented blood vessels are composed of common background in color and have no special characteristics. However, its structure is particularly messy compared to normal blood vessels. It does not have a tree-like structure formed by long blood vessels, but is mostly composed of multiple small ring structures and some finely divided small blood vessels.

Based on the above analysis, in this embodiment, the first type of mis-segmented blood vessel is determined based on the background difference on both sides of the blood vessel:

(a1) For each blood vessel, extract the feature vectors of the background on both sides of the blood vessel;

Obtain the color values of all pixels on the R, G, and B channels in the area within 10 pixels from the center line on this side and average them on each channel to obtain the feature vector of this side.

The feature vector on each side is actually a three-dimensional vector, respectively representing the color value information of the background on both sides of the blood vessel on the three channels of RGB.

(a2) Use the K-means clustering method to cluster the eigenvectors into two categories, and divide all blood vessels into two categories according to the correspondence between the eigenvectors and blood vessels. Since the probability of misjudgment is usually not too high, the obtained small categories (ie Vessels with fewer blood vessels) are the first type of mis-segmented vessels.

In this embodiment, the second type of mis-segmented blood vessels is determined based on the shape of the blood vessels:

(b1) Determine the ring structure in the fundus image demarcating the original blood vessel set.

Can construct undirected graph G=(V, E) during concrete realization, V is the set of two end points of all blood vessel centerlines, E is the set of the centerlines of all blood vessels, utilize this undirected graph G=(V, E ) to determine the ring structure.

(b2) For each annular structure, if the length of the longest blood vessel in the annular structure is less than the preset segmentation length threshold α, where α=x/60～x/45, (in this embodiment, the segmentation length threshold α =x/50, x is the lateral size of the fundus image, i.e. x=3000), then all the blood vessels in the annular structure are the second type of mis-segmented blood vessels, and further proceed as follows:

Determine the center of the ring structure, and calculate the shortest distance from the center to the blood vessel whose length is greater than or equal to α (that is, the distance from the center to the closest blood vessel whose length is greater than or equal to α), with the center as the center, the shortest distance All blood vessels in the circular area with the distance as radius are the second type of mis-segmented blood vessels.

In this embodiment, the binarization threshold is the ratio of the number of pixels that are blood vessels after the binarization process to the pixels of the entire fundus image, and the value is usually 4-20%. The larger the binarization threshold, the looser it is.

In this embodiment, six different binarization thresholds are used, which are 4%, 6%, 8%, 10%, 12% and 14%. Steps (1-1) to (1-3) are performed for each binarization threshold, corresponding to 6 global blood vessel sets respectively.

In this embodiment, the original blood vessel set obtained when the binarization threshold is 14% is shown in FIG. 4 , and the schematic diagram of the corresponding global blood vessel set is shown in FIG. 5 . It can be seen that by removing mis-segmented blood vessels, the interference caused by ring reflections caused by photographing, non-vascular step edges around the optic disc, patchy lesions, and hemorrhagic lesions can be effectively eliminated, and the accuracy of blood vessel segmentation can be improved.

In this embodiment, the optic disc positioning information is obtained by performing optic disc positioning on the fundus image. The specific process is shown in Figure 6, and the following operations are performed for each global blood vessel set:

(1-2) For each blood vessel in the current global blood vessel concentration, use a fuzzy convergence algorithm to obtain the convergence area of the blood vessel;

(1-3) Count the number of convergence areas that each pixel of the fundus image belongs to as the voting value of the pixel, and construct a voting matrix according to the voting value of each pixel, and perform mean filtering on the voting matrix, and mean filtering The size of the averaging filter used is 6×6.

Each element in the voting matrix constructed in this embodiment is in one-to-one correspondence with the pixels in the fundus image, and is the voting value of the corresponding pixel.

(1-4) According to the voting matrix after filtering, select the first n pixels with large voting values (n=3000 in this embodiment), and use the regional connection algorithm based on eight connections to obtain several connected regions for the selected n pixels. , taking the connected region with the largest area corresponding to each global blood vessel set as the final convergence region of the global blood vessel set, and judging whether there are at least l overlapping regions of the final convergence region, where l=k/2, k is a preset binary value The number of optimized thresholds, that is, l=3:

If it exists, the center coordinates of the overlapping region with the largest area are used as the optic disc positioning information;

Otherwise, the optic disc positioning information is obtained by using a specific template matching method.

In this embodiment, when selecting the first n pixels with the highest voting value, all pixels are sorted according to the voting value. In the present invention, the ranking is performed according to the voting value from large to small, and the first n pixels can be selected.

(2) Determine the main vessel according to the global vessel set with the largest (i.e. the most relaxed) binarization threshold and the optic disc positioning information, and classify the main vessel to obtain the main vessel classification information;

In this embodiment, the main blood vessels are determined by the following method:

The area within 100 pixels from the center of the optic disc is used as the adjacent area of the optic disc, and the blood vessels whose length is greater than the preset classification length threshold (in this embodiment, the preset classification length threshold is 60) in the determined optic disc adjacent area are used as the main blood vessels .

In this embodiment, the following steps are performed to classify the main vessels to obtain the classification information of the main vessels:

(2-1) Obtain the average caliber of each main vessel, and designate the main vessel with the largest average caliber as the venous vessel;

(2-2) cutting each main vessel into several segments to obtain corresponding main vessel slices;

In this embodiment, each pixel is a slice along the central line of the main blood vessel during segmentation, and then each main blood vessel is cut (ie blood vessel slice) into several segments.

(2-3) Obtain color information of 5 pixels along the vessel center of the main vessel slice to both sides, and use the mean value of the color information of all pixels as the feature vector of the main vessel slice;

In this embodiment, the RGB value and the HSL value of the pixel point, that is, the obtained feature vector is a 6-dimensional vector, and each dimension corresponds to the color value of the pixel point on the R, G, B and H, S, L channels.

Then, based on the eigenvectors, K-means clustering method is used to cluster all the main vessel slices into two categories, and the category corresponding to the main vessel slices corresponding to the venous vessels is regarded as the venous vessel category, and the other category is regarded as the arterial vessel category.

(2-4) For each main vessel, the class in which more main vessel slices belong is used as the classification result of the main vessel.

For example, for any main vessel, A% of the corresponding main vessel slices are in the arterial category, and B% are in the venous category. If A is greater than B, the vessel is considered to be an arterial vessel, and if A is less than B, it is considered to be an arterial vessel. The blood vessel is a venous blood vessel, otherwise, it is arbitrarily designated.

(3) Using the main vessel classification information, the SAT-based breadth search algorithm is used to classify the vessels in the global vessel collection to obtain the global classification information.

For the breadth search of the global blood vessel set, the breadth search algorithm based on SAT uses three constraints to transfer the blood vessel category during the search process: the two blood vessels that cross each other are marked as two types of blood vessels; the three-fork blood vessels are marked as one type of blood vessels ; If the sum of the included angles between one of the three-chambered blood vessels and the remaining two blood vessels is less than or equal to 270 degrees, it is judged as a three-chambered structure of constraint 2, otherwise no judgment is made.

Fig. 7 is the global classification information obtained by using the breadth search algorithm based on SAT in this embodiment, it can be seen that the number of blood vessels that are not classified (that is, there is still no classification information after classification) is relatively small, and the three-dimensional error caused by blood vessel segmentation is greatly reduced. Situations where vessel classification in bifurcation structures cannot be performed.

Unless otherwise specified, the boxes with rounded corners in all the flow charts in this embodiment represent the results obtained, and the rectangles with square corners represent operations.

The above-mentioned specific embodiments have described the technical solutions and beneficial effects of the present invention in detail. It should be understood that the above-mentioned are only the most preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, supplements and equivalent replacements made within the scope shall be included in the protection scope of the present invention.

Claims (8)

1. the arteriovenous retinal vessel sorting technique based on the eye fundus image of breadth first search method, first overall blood vessel collection and the optic disc locating information of eye fundus image is obtained, described overall blood vessel collection is the set of all blood vessels in described eye fundus image, described optic disc locating information comprises the optic disc center of described eye fundus image, then the classification of arteriovenous retinal vessel is carried out according to described overall blood vessel collection and optic disc locating information, it is characterized in that, during classification, carry out following steps:

(1) determine main blood vessel according to described overall blood vessel collection and optic disc locating information, and classification is carried out to main blood vessel obtain main blood vessel classified information;

(2) the main blood vessel classified information described in utilization adopts the breadth first search method based on SAT to carry out classification to the blood vessel that described overall blood vessel is concentrated and obtains global classification information.

2., if claim 1 is based on the arteriovenous retinal vessel sorting technique of the eye fundus image of breadth first search method, it is characterized in that, described step (1) determines main blood vessel by the following method:

Using the region within several pixels of distance optic disc center as optic disc adjacent domain, in described optic disc adjacent domain, length is greater than the blood vessel of default classification length threshold as main blood vessel.

3., if claim 1 is based on the arteriovenous retinal vessel sorting technique of the eye fundus image of breadth first search method, it is characterized in that, described classification length threshold is 50 ~ 65.

4. if claim 2 is based on the arteriovenous retinal vessel sorting technique of the eye fundus image of breadth first search method, it is characterized in that, described step (1) is carried out classification to main blood vessel as follows and is obtained main blood vessel classified information:

(1-1) obtain the average caliber of each main blood vessel, the main blood vessel of specifying average caliber maximum is vein blood vessel;

(1-2) each main blood vessel is cut into some fragments, obtains corresponding main Vascular Slice;

(1-3) characteristic vector of each main Vascular Slice is extracted, and adopting clustering procedure based on described characteristic vector, described main Vascular Slice to be gathered be two classes, and using by the class at main Vascular Slice place corresponding for vein blood vessel as vein blood vessel, another kind of as arteries;

(1-4) for each main blood vessel, using the class at more main Vascular Slice place as the classification results of this main blood vessel.

5. if claim 4 is based on the arteriovenous retinal vessel sorting technique of the eye fundus image of breadth first search method, it is characterized in that, adopting K means Method described main Vascular Slice to be gathered in described step (1-3) is two classes.

6., if claim 4 is based on the arteriovenous retinal vessel sorting technique of the eye fundus image of breadth first search method, it is characterized in that, in described step (1-3), extract the characteristic vector of each main Vascular Slice by the following method:

Obtain the colouring information apart from all pixels in the region within several pixels of blood vessel center of main Vascular Slice, and using the average of the colouring information of all pixels in this region as the characteristic vector of this main Vascular Slice.

7. if claim 6 is based on the arteriovenous retinal vessel sorting technique of the eye fundus image of breadth first search method, it is characterized in that, when extracting the characteristic vector of each main Vascular Slice, obtain the colouring information apart from area pixel point within blood vessel center 5 ~ 8 pixels of main Vascular Slice.

8. if claim 7 is based on the arteriovenous retinal vessel sorting technique of the eye fundus image of breadth first search method, it is characterized in that, described colouring information comprises rgb value and the HSL value of this sampled point.